Qwen4-Max
Points and comments are a snapshot, not live.
Qwen4-Max is a 2.4-trillion-parameter AI model with a 1M-token context window.
Qwen4-Max, built on a new sparse MoE architecture (71B activated per token, 328B N-gram Embedding parameters), introduces GDN + Qwen Sparse Attention and Gated Residual branches. It shows performance gains over Qwen3.8-Max: Work360 score 52.69 vs 47.74, token usage drops 24-49%, average cost per task down 62% to $1.67. It handles 3D modeling, game development, automotive manufacturing, video production, finance, law, and healthcare tasks. In cloud resource scheduling, it reduced annotation costs 10.42% and improved solving speed 14.61x.
It powers 24/7 agents with Live Delivery for iterative task execution and user control.
What commenters are saying
Commenters are cautiously optimistic but raise concerns. Some praise the focus on efficiency and token savings, hoping for smaller Qwen4 Flash/27B versions that fit consumer hardware (e.g., 96GB Mac Studio). Others complain about Qwen3.8 speed (30 tok/s vs 80) and limited third-party providers, noting the open-source version lacks vision support. Several dismiss the Artificial Analysis benchmark as manipulated or irrelevant, citing weight changes and non-native harnesses. Speculation that Anthropic blocked distillation of Opus, and claims Qwen4-Max cannot compete with frontier models.